🌾 Does Better Data Make Better Markets?

1. The Question

Agricultural markets have always operated under uncertainty. A farmer planting tomatoes in March doesn't know what the weather will be in July, what diseases might strike, or what the spot price will be at harvest. An investor prepaying for a share of that harvest faces the same fog — plus not knowing what other investors are doing.

Precision agriculture technologies — soil sensors, satellite imagery, drone-based crop monitoring — are generating unprecedented data about what's actually happening on farms. The question is:

We built an agent-based simulation to find out. The answer turns out to be surprising: knowing what the market thinks matters more than knowing what the crops are doing.

2. The Model: How It Works

The simulation models a market where investors prepay for fractions of a farm's future produce — like agricultural futures, or community-supported agriculture (CSA) shares:

One simulation step equals one week. A typical run covers 2 growing seasons (65+ weeks). Weather drives crop growth, crop growth drives expected yields, yields drive contract valuations, valuations drive bidding, bidding drives prices. Everything is connected.

3. Information Tiers: The Controlled Variable

The central experiment varies how much information agents have:

4. The Mathematics Under the Hood

Crop Growth — GDD-Based Logistic Model

Biomass follows a logistic curve driven by Growing Degree Days:

B(t) = B_max / (1 + exp(-k × (GDD(t) - GDD_mid)))

Where:
  GDD accumulates weekly as max(0, T_avg - T_base) × 7
  Yield = Biomass × Harvest_Index × Stress_Multiplier
  Stress = (1-water_stress×0.6) × (1-temp_stress×0.5) × (1-disease×0.4)

Weather — Seasonal Stochastic Process

T(week) = T_mean + T_amp × sin(2π(week - φ)/52) + N(0, σ)
Rain ~ Gamma(shape, scale × seasonal_factor)
Climate modes: normal, El Niño (+1.5°C, -30% rain), La Niña (-1°C, +30% rain)

Spot Prices — Ornstein-Uhlenbeck

dP = θ(μ - P)dt + σdW    (mean-reverting commodity prices)

Livestock — Gompertz Growth

W(t) = W_max × exp(-b × exp(-c × t))    (asymptotic weight gain)

5. Agent Types and Strategies

Investor Strategies

• Speculator (25%) — Aggressive bidding (1.10× multiplier), low risk aversion, flips contracts on secondary market for profit
• Conservative (30%) — Careful valuation (0.95×), high risk aversion, holds to harvest for steady returns
• Momentum (15%) — Follows price trends — bids more aggressively when prices are rising
• Contrarian (15%) — Counter-trend — buys aggressively when prices are falling (bargain hunting)
• Value (15%) — Fundamentals-based — applies a premium when prices are below long-term average

Bank Lending Behaviors

Banks evaluate loans using a credit score derived from the borrower's capital, utilization rate, and repayment history. They also adjust rates based on what competing banks are offering.

6. Results: What We Found

We ran 50 Monte Carlo simulations per information tier, holding all other parameters constant. Here are the key findings:

The Information Hierarchy

Our results suggest a clear hierarchy of information value:

Market consensus (public price index)
  > Competitor awareness (who else is bidding)
    > Farm-level telemetry (precision ag data)
      > Personal history (can be harmful!)

Key Metrics by Tier

7. Seeing Causation: The Step-Through Playback

The simulation features an interactive step-through mode where you can scrub through every week and observe the causal chain:

You can use the ◀ Back and Fwd ▶ buttons, or drag the week slider, to move through time and watch how:

• A drought week reduces soil moisture → crops stress → expected yields drop → investors bid less → prices fall
• A bank loan gives a cash-strapped farmer operating capital → they list more plots → more supply → prices stabilize
• Heavy rainfall triggers disease → disease reduces yield forecasts → precision-tier investors adjust bids downward while blind-tier investors don't notice

8. Implications for Precision Agriculture

For Market Design

If the goal is market fairness and efficiency, investing in public market transparency infrastructure (real-time price indices, open contract databases) may be more impactful than subsidizing private farm-level monitoring technology. A simple published average price does more for Gini reduction than sophisticated crop sensors.

For Platform Design

Market platforms should prominently display public price indices rather than letting participants' interfaces show only their personal transaction history. The "local" tier's poor performance shows that personal anchoring actively harms market efficiency.

For Investors

More information doesn't necessarily mean more profit. The "full" information tier (market + competitors) produced the most competitive environment and the lowest individual returns. Information is an arms race — when everyone has it, the edge disappears.

For Farmers

When markets are more transparent (market/full tiers), contract prices are more competitive — good for farmers who receive higher bids. But it also means more investors participating, which can create auction pressure and volatile bidding patterns.

9. Try It Yourself

The full simulation is available as an interactive Streamlit app:

Three Modes

1. Single Run (Step-Through) — Run a simulation, then scrub through every week. See the causal chain, watch crops grow, track money flow, inspect every trade.
2. Monte Carlo Experiment — Run 50+ simulations per information tier. Compare box plots, effect sizes (Cohen's d), strategy × tier interactions. Export as CSV.
3. Crop Growth Explorer — Adjust temperature, rainfall, and irrigation interactively. See how each crop responds. Compare all crops under the same conditions.

Quick Start

# Clone and run locally
git clone https://huggingface.co/spaces/rdjarbeng/agrimarket-abm
cd agrimarket-abm
pip install mesa networkx pandas numpy streamlit plotly scipy
streamlit run app.py

Run Monte Carlo from Python

from metrics import run_monte_carlo, compute_effect_sizes

df = run_monte_carlo(
    info_levels=("blind", "local", "market", "full", "precision"),
    n_runs=50,
    num_investors=20,
    num_farmers=2,
    num_banks=2,
    season_length_weeks=30,
    num_seasons=2,
)

effects = compute_effect_sizes(df, baseline_tier="blind")
for tier, metrics in effects.items():
    sig = [m for m, d in metrics.items() if d["significant"]]
    print(f"{tier}: significant effects on {sig}")

10. What's Next

Planned Features

• Factory buyers — Ketchup factory for tomatoes, chocolate factory for cocoa. Guaranteed demand at fair price, but with capacity limits that force surplus to the open market
• Post-harvest losses — Model spoilage, storage costs, and transport losses. Farmers react by adjusting planted area
• Adaptive lot sizing — Farmers expand plots when demand is high and banks are lending; contract when post-harvest losses are high or investment dries up
• Cash-strapped scenarios — Test what happens when investors start with limited capital, or when farmers need loans just to plant
• Reinforcement learning agents — Replace rule-based strategies with agents that learn optimal bidding through experience
• Multi-region trade — Markets across different climate zones with transport costs

Research Extensions

• Integration with real crop models (APSIM, DSSAT) for agronomic realism
• Validation against historical commodity market data
• Policy experiments: crop insurance, price floors, information subsidies
• Network effects: agents sharing information through social ties

Architecture

crop_models.py  — CropGrowthState, LivestockGrowthState, WeatherEngine, SpotPriceEngine
agents.py       — Plot, Farmer, Investor, Bank, PriceManager
model.py        — AgriMarketModel (Mesa 3.5 ABM orchestrator)
metrics.py      — compute_run_metrics, run_monte_carlo, compute_effect_sizes
app.py          — Streamlit dashboard with step-through playback

Built with Mesa 3.5 (agent-based modeling), Streamlit (dashboard), Plotly (visualization), and Python 3.12.